Sang-Bo Park

The latest convolutional neural network (CNN) models for object detection include complex layered connections to process inference data. Each layer utilizes different types of kernel modes, so the hardware needs to support all kernel modes at an optimized speed. In this paper, we propose a high-speed and optimized CNN accelerator with flexible diagonal cyclic arrays (FDCA) that supports the acceleration of CNN networks with various kernel sizes and significantly reduces the time required for inference processing. The accelerator uses four FDCAs to simultaneously calculate 16 input channels and 8 output channels. Each FDCA features a 4 × 8 systolic array that contains a 3 × 3 processing element (PE) array and is designed to handle the most commonly used kernel sizes. To evaluate the proposed CNN accelerator, we mapped the widely used YOLOv5 CNN model and evaluated the performance of its implementation on the Zynq UltraScale+ MPSoC ZCU102 FPGA. The design consumes 249,357 logic cells, 2304 DSP blocks, and only 567 KB BRAM. In our evaluation, the YOLOv5n model achieves an accuracy of 43.1% (mAP@0.5). A prototype accelerator has been implemented using Samsung’s 14 nm CMOS technology. It achieves 1.075 TOPS, a peak performance with a 400 MHz clock frequency.

The rapid advancement in AI requires efficient accelerators for training on edge devices, which often face challenges related to the high hardware costs of floating-point arithmetic operations. To tackle these problems, efficient floating-point formats inspired by block floating-point (BFP), such as Microsoft Floating Point (MSFP) and FlexBlock (FB), are emerging. However, they have limited dynamic range and precision for the smaller magnitude values within a block due to the shared exponent. This limits the BFP’s ability to train deep neural networks (DNNs) with diverse datasets. This paper introduces the hybrid precision (HPFP) selection algorithms, designed to systematically reduce precision and implement hybrid precision strategies, thereby balancing layer-wise arithmetic operations and data path precision to address the shortcomings of traditional floating-point formats. Reducing the data bit width with HPFP allows more read/write operations from memory per cycle, thereby decreasing off-chip data access and the size of on-chip memories. Unlike traditional reduced precision formats that use BFP for calculating partial sums and accumulating those partial sums in 32-bit Floating Point (FP32), HPFP leads to significant hardware savings by performing all multiply and accumulate operations in reduced floating-point format. For evaluation, two training accelerators for the YOLOv2-Tiny model were developed, employing distinct mixed precision strategies, and their performance was benchmarked against an accelerator utilizing a conventional brain floating point of 16 bits (Bfloat16). The HPFP selection, employing 10 bits for the data path of all layers and for the arithmetic of layers requiring low precision, along with 12 bits for layers requiring higher precision, results in a 49.4% reduction in energy consumption and a 37.5% decrease in memory access. This is achieved with only a marginal mean Average Precision (mAP) degradation of 0.8% when compared to an accelerator based on Bfloat16. This comparison demonstrates that the proposed accelerator based on HPFP can be an efficient approach to designing compact and low-power accelerators without sacrificing accuracy.